2170 - Development of a Tissue Segmentation-Mediated Synthetic CT Method for Improved Gastrointestinal Gas Cavity Definition for Online Adaptive MRI-Guided Radiation Therapy

Purpose/Objective(s): Deep-learning-based synthetic computed tomography (sCT) images can provide electron density information for dose calculation for online adaptive magnetic resonance imaging (MRI)-guided radiotherapy. Single-stage sCT model training requires deformably registered MRI/CT images, resulting in inconsistent gastrointestinal (GI) and gas cavity definition in sCT images. Manual correction of gas cavity definition errors in sCT is time consuming, and during this process, gas cavities can migrate potentially altering dosimetry for online adaptive plans. In this study, we propose a two-stage sCT model which improves gas cavity definition compared to single-stage sCT methods.

Materials/Methods: SCT images were generated using two generative adversarial network (GAN) models. The first model converts an MRI image to a segmentation map and the second model converts a segmentation map to a sCT image. The MRI-to-segmentation map was created using paired MRI/segmentation maps using the CycleGAN model. The sCT image was generated using segmentation map inputs with a separately trained conditional GAN model, pix2pix. Model validation and hyperparameter tuning was done with five-fold cross-validation with paired MRI/segmentation maps and segmentation maps/CT images from 25 patients.

A multi-stage thresholding method was used to define GI gas cavities. First, GI contours were created with artificial intelligence (AI) autocontouring and manually verified. Second, gas cavities were delineated with 20% intensity threshold within the AI-generated GI contour of each patient.

We quantitatively compare gas cavity definition for our two-stage sCT model to the standard single-stage MRI/CT model. Mean absolute errors (MAE) values in Hounsfield units (HU) were computed to evaluate global sCT accuracy. DICE coefficients were computed to evaluate the gas cavity definition accuracy in segmentation maps.

Results: The global MAE for the two-stage method and the standard single-stage method were 84HU±15HU and 122HU±19HU respectively. The gas cavity DICE for the two-stage method and the single-stage method were 0.67 ± 0.08 and 0.51 ± 0.08 respectively.

Conclusion: The proposed two-stage model improves GI gas cavity definition in sCT images compared to the standard single-stage model. Future work will include prospectively evaluating the improvements in dosimetry and workflow efficiency from using this method for online adaptive MRI-guided RT.